The Teaching Machine – Past, Present and Future

Robert N. Bilyk
President, LodeStar Learning Corporation

I’ve been reading Teaching Machines, The History of Personalized Learning by Audrey Watters¹, who may be best known for her “Hack Education” blog.  Teaching Machines is a great read.   It was published in 2021.  Since then, Audrey Watters posted her last on ‘Hack Education’ and, according to her post, no longer writes about Tech Education. 

The Past

A good part of Teaching Machines is dedicated to B.F. Skinner’s Teaching Machine.  By today’s standards, it’s a crude mechanical device.  The early versions were wood boxes that displayed questions through a small window and required students to write answers on a paper strip.  Once students submitted their answers, they pulled a lever to display the answer and advance the paper strip.  The questions and answers were printed on a paper disk partitioned into pie wedges.

For fun, I simulated an early prototype of B.F. Skinner’s Teaching Machine.  I used our LodeStar Authoring tool to simulate how the early machine worked.

An early prototype of B.F. Skinner’s Teaching Machine

See the simulation at https://lodestarlearning.github.io/Teaching-Machine/index.htm

Films from that era (the 50s) conjure up rows of students working on the machines – a setting not unlike a factory floor with row upon row of workers at their sewing machines.

And yet, as Watters writes, the impetus – the motivation – behind the design of these teaching machines was largely, in the opinion of their inventors, altruistic and intended for the benefit of the student.

Watters recalls how Skinner visited his daughter’s 4^th grade math class.  The teacher wrote math problems on the board.  Some students finished quickly and squirmed in their seats as they waited for the class to progress.  Other students struggled to complete the problem and perhaps never successfully finished.

Fast forward thirty years.  I was a high school language arts teacher in the 80s.  One of my classes consisted of students who could not read or write very well.  I’ve spoken about this a dozen times.  The underlying reasons for their inability to read or write very well had to do with who they were:  some students were newly immigrated from non-English speaking countries; other students had learning disabilities; and others hated school.  I was handed one curriculum that treated the class as one homogenous body.  I really needed materials and strategies that addressed the needs of at least three populations.  I needed some form of teaching machine.

Skinner, in response to his daughter’s class, decried that “something must be done”.  I approached our Instructional Materials Center Coordinator and asked if, in any way, the Apple IIe machine in the library could help me ‘individualize’ the instruction.  For Skinner, the realization led to his idea of a teaching machine.  For me, I began a career-long journey into the world of technology-assisted personalized learning.

Minnesota Educational Computing Consortium’s (MECC) Oregon Trail on an Apple II.

From time to time, I’ve challenged myself to take a hard look at the principles that drove my pursuit to help students with technology.  As a starting point, I was interested in the following ideas:

For the student:

• Go at your own pace, where the learning objective is constant, but time is variable.
• Get quick feedback.
• Have fun.
• Pursue lesson branches based on your interest, choices and performance.
• Be challenged.

For the instructor:

• Free up time for individualized help.
• Get information on what questions students are consistently getting wrong. 
• Inform future lesson plans.

I used computer programs on the Apple IIe, but I wanted to write my own.  Several years later, I was headed home on a dark country road, listening to Minnesota Public Radio.   I was listening to a story about interactive video discs – and I became smitten.  At this point in my career, I worked at a college.  I asked my dean (who had a background in instructional systems), what I needed to learn to program interactive discs.  He answered, “Learn C”.  And so I began my study of both learning theory and computer programming and camped out at their crossroads.

At the University of Minnesota, I studied curriculum and instruction.  At Saint Paul College, where I worked, I took classes from programmers who made their living at Unisys, Honeywell and other industries.  As I soon discovered, my dean’s advice was both bad and good.  It was bad because, by the 90s, one did not need to learn a low-level programming language, like C, to control a video disc player.  The TenCore Language Authoring System could do that with a much higher level language (thus easier).   It was good because I became proficient at programming and could extend authoring systems with lower-level code. In short, I was never limited by the authoring tool, only by my imagination. Later I worked on teams to design a content management system, several instructional products and two authoring systems.

The Tencore Language Authoring System (LAS)
was a TUTOR (PLATO) derivative developed by Paul Tenczar 

The Interactive Videodisc was a laser– readable random-access disc that contained both audio and analog video. The videodisc’s full motion video was displayed on a computer with the help of a video overlay card that superimposed the analog video on a digital monitor.

Authoring systems like TenCore, Authorware and Toolbook enabled designers to program the computer to send commands to the player via serial communications.

In the early 90s, the University of Minnesota introduced me to Course of Action, which was later renamed Authorware.  Authorware was a game-changer.  With this system, I could finally efficiently realize all the ideas I had started out with and that now took the form of:

• Individualized instruction
• Immediate feedback, or strategically delayed feedback, based on the design.
• Fun interaction with analog video, high-fidelity sound, dithered graphics (256 color displays), three-dimensional animation (yes, in the early 90s), simulations, games….
• Branched, non-linear instruction.
• Data collection and reporting. (The installed computer program could gather data and, using an early standard from the aviation industry (AICC), report it to the instructor.)

It was a great start and a step toward the present. 

Authorware, designed by Dr. Michael Allen and his amazing team, changed everything. The work shifted from arcane computer commands to the design of learning experiences. Authors dragged icons onto a flowline that commanded the computer to present graphics and text, animate, play audio and video, branch instruction and execute simple scripts.

The Present

The promise of personalized instruction is partially realized today.  Our current teaching machines are programmed by smart authoring systems that can present and control media, animate, branch instruction, simulate real-world conditions in various degrees of fidelity, and report student performance through standards like SCORM, xAPI and CMI5.  We also have virtual reality, which can immerse learners in some ‘real-world’ or imaginary context, and augmented reality, which can add a digital layer over one’s interaction with the physical world.

We also benefit from gamification, adaptive learning, learning experience platforms, learning management systems, microlearning, digital story-telling, three-dimensional graphics, and 360 imagery and video.  In short, Teaching Machines have come a long way from questions displayed in a tiny window.

And yet there is still something significantly missing that would have benefited both Skinner’s daughter and my high school students…or any learner.

The Future

When it comes to the future, I’m on thin ice.  I offer this up more as an invitation to hear from others who have greater insights than I. There is so much happening in academia, industry and in academic-commercial partnerships that I’m just not aware of.  But I’ll briefly touch on what has been nagging me for years – and return to this theme in a later post with, perhaps, your contributions.

The problem, in my view, is in the finiteness of our present- day teaching machines.  As an example, years ago, my colleagues put out a math series on interactive video disc, then later on CDROM, and then later on the web.  The lessons were well presented and followed by numerous exercises in various topics like college-level algebra.

I observed students working through the math programs in Saint Paul College’s Instructional Technology Center.  If students didn’t ‘get it’ from the presentation or the exercise feedback, they had little recourse.  Fortunately, in some settings, a qualified instructor would then take over and help the student with a different tactic.  I observed settings, however, where similar programs were being used but with underqualified tutors.  If the students didn’t comprehend the lesson, that was the end of the line.  The lab monitor was unable to help them.

A lot of topics rely on a student’s prior knowledge, which may or may not be present.  In short, prerequisite skills. 

Several years ago, a fleet of‘adaptive learning’ platforms appeared on the market that proposed to remedy this short-coming.  (examples of adaptive learning systems include CogBooks, Acrobatiq, Knewton).  A lot of investment and effort went into assessing student’s performance and confidence with the subject matter and providing alternative instruction.  Some of the platforms catalogued open educational resources and used, what they called, semantic engines to match students with open resources to help them.

Adaptive learning systems are software platforms that optimize the content to adjust for the learner’s goals and current state of knowledge. They are designed to address challenges such as slow pace, difficulty, lack of practice, and insufficient mastery of content. In traditional e-learning courses, students linearly follow the path an instructor creates. They watch videos, read articles, take quizzes, and practice interactive modules often in a predetermined order. An adaptive learning system contains the same types of material but the order, pacing, and content may change for each learner. The system decides which content to show the learner based on the learner’s goal, performance and confidence level.  If the system determines the current path is too easy for the learner, it can branch to more challenging material. If it finds out the current path is too difficult, it may intervene and review prerequisite content, reduce the challenge, or slow down the pace. 

Most of these systems are proprietary and operate through subscription or a pay wall.  Presumably, students who use them benefit from them but the modern teaching machine (i.e., authoring systems, Learning management systems, learner experience platforms) needs to tap into the same technology and benefit from open systems.

Work is being done to open-source systems that use such approaches as Bayesian Knowledge Tracing.  Bayesian Knowledge Tracing is used in a number of Cognitive Tutoring systems.  This approach develops a model of a student’s knowledge in a given domain and constantly updates that model based on the student’s performance.  Probability of skill mastery can be computed from the student’s current knowledge and the proposed learning material.  Efforts are underway to use these systems to link learning management systems with open content repositories.

 A possibility in the future is that the entire web becomes the new teaching machine.  In the past, several concepts and specifications rose and fell in popularity.   We may see their renaissance.  One concept was Tim Berners-Lee’s Semantic web.

The goal of the semantic web as it relates to education is to make content on the web machine readable – and therefore categorizable and discoverable.  To support the description of educational resources several specifications exist including the Resource Description Framework, and the Web Ontology Language. 

A recently published article titled ‘Investigating the potential of the semantic web for education: Exploring Wikidata as a learning platform’²  explores the application of the semantic web in academic pursuits. 

In short, if one could accurately describe educational content (which may reside in an open educational repository), then authoring tools, Learning Management Systems, adaptive systems, etc. could discover and propose the content to a student who is struggling to achieve an objective.

The key is in an accurate and meaningful description so that its appropriateness can be evaluated for a given student and a given objective.

In the past, we haven’t done well with descriptions, ontologies, taxonomies, metadata and the like.  (By we, I mean the teaching profession.)  Even the concept of the re-usable learning object (RLO) fell to criticism.  A learning object was described by IEEE Learning Object Metadata specification.  The LOM standard defined the attributes required to describe learning content. Unfortunately, learning objects were created without adequate descriptions (after all, what was the point?) and their reusability was called into question.  In general, it’s possible that content authors don’t have the knowledge or the means to properly describe their content in a way that is meaningful to authoring, tutoring or adaptive systems.

The game-changer will probably be machine learning.  Blogs have been filled with examples of how artificial intelligence has supported content generation.  Another important use of AI will be to scan and describe open learning content.

For fun, I asked ChatGPT to parse some educational content on an electrical circuit and then provide me with a description that followed both the Resource Description Framework and then IEEE Learning Object Metadata.   I then asked ChatGPT to come up with its own machine-readable schema that included keywords and reading level.   The simple schema (markup in XML) is shown in the appendix.

The future Teaching Machine may be the semantic World Wide Web

Conclusion:

In my view, to personalize learning, the future Teaching Machine needs to tap into the World Wide Web as a resource to shore up prerequisite knowledge needed to achieve an educational objective.  The systems can’t be proprietary.  They must be open and available to a variety of platforms through an application programming interface (API).  The platforms include authoring systems like LodeStar eLearning authoring system, Captivate, and Storyline as well as learning experience platforms, adaptive learning systems, cognitive tutors and learning management systems. 

As importantly, the resources should remain in their repositories, but still report performance to a learning management system or learning experience platform.  That is now possible through standards like CMI5, but — as I’ve lamented in the past — few learning management systems support CMI5.

The teaching machine began with limited content printed on paper disks.  It progressed to include all of the affordances of the modern computer.  Its future promise might lie in open standards that connect learner needs to appropriate content scattered across the globe.

References

¹Watters, A. (2021). Teaching machines the history of personalized learning. The MIT Press.

² Evenstein Sigalov, S., Nachmias, R. Investigating the potential of the semantic web for education: Exploring Wikidata as a learning platform. Educ Inf Technol 28, 12565–12614 (2023). https://doi.org/10.1007/s10639-023-11664-1

Appendix

ChatGPT generated XML schema for a learning resource on electrical fundamentals.

<educational_resource>

    <title>Understanding Electrical Fundamentals: Voltage, Current, and Resistance</title>

    <keywords>

        <keyword>Voltage</keyword>

        <keyword>Current</keyword>

        <keyword>Resistance</keyword>

        <keyword>Electricity</keyword>

        <keyword>Electrical Power</keyword>

        <keyword>Ohm’s Law</keyword>

        <keyword>Circulatory System</keyword>

        <keyword>Electrical Concepts</keyword>

    </keywords>

    <reading_level>High School / Early College</reading_level>

    <content>

        Just as your heart produces the pressure to make blood circulate, a battery or generator produces the pressure or force to push electrons around a circuit. …..(the remaining content omitted for the sake of brevity)

    </content>

</educational_resource>

Beyond Presentations and Puzzles … to Problems

Introduction

The thrill of learner experience design is in finding the tools, techniques and strategies to engage learners with designs that go well beyond presentations and puzzles.  Often learners must solve problems to achieve a performance or academic goal and it’s in the nature of the problem that we find the best-suited strategies.

I use the term ‘problem’ very broadly to encompass many different things.  The late David H. Jonassen presented a typology of problems in his work, ‘Toward a design theory of problem solving’¹.  He wrote that ‘learning to solve problems is too seldom required in formal educational settings.’  Even outside of formal education, we’re often asked to create presentations with assessments that cover some area of compliance whether it be data security, diversity in the workplace, or health and safety.  But often we’re challenged with a training problem that can’t be addressed without engaging learners with the problem.

David H. Jonassen published many books on constructivism, problem-solving and learning with technology.

The Challenge

Our challenge is to develop the right thing that gets the best results with the least expenditure of time and money.  We need strategies that are well-suited to the problem. Seasoned designers have a wide repertoire of strategies, templates, and models that can be used given a level of learning, type of learning, and type of problem.  However, this industry sees new entrants every year² and the casual developer (an instructor, for example) may not have uncovered all of their options or know when to apply them.

What are the activity types that provide a situation, a problem to solve, and an activity that elicits student performance and constructive feedback?  What are the activities that help designers and instructors engage students at a level that involves challenge, activity and feedback.  More generally, what activities promote higher order thinking that allow us to integrate component skills into a coordinated response that enables us to complete tasks, analyze data, solve problems, and create things?

We have limited time and money.  The quick solution to any higher order learning objective is to simply talk about it, present on it, show a video, hear from an expert, ask some questions, and then, forgive me, throw the learners in the deep end.  Moreover, if we take the time to create an activity that makes the learner do things we may ask ourselves:  was the result commensurate with the time and effort?   In other words, given a learning objective and educational goal, the online instructor or trainer might wonder about the return of investment for a given strategy.  For example, an instructor may ask ‘Is it worth my time to create an activity rather than a multiple choice test?’   And what would that activity be?

Our industry doesn’t make it easy.  It turns out that the industry has many names for things.  There is no standard taxonomy with precise attributes.  And yet we can start with a few widely recognizable terms:  Interactive case study.  Decision-making scenario.  ShortSims³.  Simulation. Branching interaction.  Interactive Non-Fiction. 

Matching Strategies to the Problem

Our first clue that we need more than a presentation is that we’re asking learners to do more than just recall.  We recognize a need to engage learners’ ability to analyze, apply, synthesize, evaluate, make decisions, create things – all of which requires thinking beyond simple recall. Of course, the learner may need to recall the right fact, principle, concept, or rule that is useful in a given situation.  But it is the application of that information that is the key. We recognize that we need a story, a context, possibly varying levels of difficulty, and possibly a challenge in the form of incomplete information, confusion and stress. 

So what’s the solution?  The solution is suggested by the problem.

We can be guided by the nature of the problem that we want learners to solve or engage in.  M. David Merrill tells us that ‘effective instruction is problem-centered’.  Robert Gagné asserted that the point of education was to help learners become better problem solvers. Michael Allen emphasizes the importance of context (‘a problem-solving environment’) and challenge. Richard Mayer wrote that ‘a major challenge of education is improving students’ minds–a goal that is reflected in people being able to solve novel problems they encounter’. 

In ‘Toward a Design Theory of Problem Solving’, David H. Jonassen creates a typology of problems – a useful tool that might help in the selection of instructional activities to promote problem solving.

Categorizing Problems

David Jonassen gave us a way to categorize problems based on problem variation.  Does the problem have multiple solutions or only one?  How many steps are included in the solution?  Does the problem solution draw from multiple principles, concepts, or even disciplines?  Does it require research beyond the content offered in a program of study?  Is the problem domain specific – can it only be solved by learners who are schooled in a specific area of study — or is it general?  Do learners need varying stages of support or scaffolding as they engage in the problem?

If we’re tackling one aspect of the problem, we can think of a key principle from ‘How Learning Works’ ³. Loosely, learners need practice on component skills and then practice with integrating them.  Is our learning objective focused on one component skill or does it involve integration of multiple skills?   If the component skill is simple enough, we might prescribe an activity like a word problem or matching type of exercise.  If we engage students in a synthesis of skills, we’re going to need something more – something that has story (context) and multiple challenges. Jonassen recommends embedding instruction in some authentic context. Context is also a key element to the CCAF model.

Now we’re in the territory of ShortSims (generically, simple sims), decision making scenarios, interactive non-fiction, interactive case studies and other forms of interactions that place learners in realistic settings, challenge them, and make them do things.

A Document Management System was simulated with the LodeStar eLearning authoring tool that challenged new employees to find key documents based on varying search criteria

One way of rationalizing the design and development time relates to the complexity of thinking and acting …and the complexity of the problem.

In a ShortSim the learner is presented with choices.  The choices can cause the narrative to branch or simply advance the narrative in a linear progression.  Making better choices might cause the learner to recall past learning, read and analyze information while looking for clues.  The learner might wonder ‘Is this an instance of, for example, the unconscious bias that I’ve been told about?  If I make the wrong choice, am I compromising data security, risking my safety, … making my supervisor unhappy?’

A ShortSim can engage the learner in some pretty advanced stuff with the clever use of graphics, text and choices.

Clark Aldrich wrote a How-To book, ShortSims,
that walks designers through the process of planning and producing
a simple decision-making simulation. Full reference below

Further down the continuum of simulations, our learning objective might require the learner to make decisions in an environment that includes many variables.  The underlying model might not be a finite set of choices but a complex interplay of multiple variables.  The activity might require a coordinated, intelligent response to an infectious disease outbreak, for example.  The underlying model might be the SEIR model, which represents a population that is susceptible to a specific disease, exposed, infected and partly recovered.   The learning program might vary the number of treatment days, incubation period, contacts per day, and fatality rate.  The learner might need to observe the data and then prescribe a response.  The computer environment can present one scenario from a nearly infinite number of possible scenarios. 

The choice of one strategy over the other can relate to the complexity of the problem we are wanting learners to solve.  Jonassen helps us here. The level of realism and variability of the problem will require increasingly more sophisticated (and time-consuming) designs.

Logical problems that require students to recall information might suggest a word game, a crossword puzzle, or a multiple-choice quiz.

Algorithmic problems require learners to progress through a series of procedural steps and this might suggest a ShortSim or decision-making scenario. 

Story problems with embedded procedures might suggest word problems or interactive fiction or non-fiction that houses word problems in a longer narrative.

Rule-Using Problems and Decision-Making problems that involve defined procedures and methods can be handled with ShortSims, other forms of interactive branching and decision-making scenarios, and role-playing.

Trouble-shooting problems can begin with interactive trouble-shooting guides or low fidelity simulations but when the problem state has variation and complexity then a virtual lab or full-blown simulation might be needed that can a) generate variable states, b) represent those states to the learner in the form of dials, instruments, changes in the environment, changes in characters, etc., c) elicit learner response and d) and provide feedback in the form of dialog, pop-ups, and/or changes in the environment.

Case-analysis problems can be presented through problem-based learning scenarios, historical case analyses, interactive case studies in which the learner is presented with case descriptions and data and asked to accomplish a goal.

Additional forms of problems in Jonassen’s typology include design problems and dilemmas. The types of problems can certainly be expanded to include additional categories of problems: for example, strategic performance problems that involve business simulations and strategy games. And the list goes on …. all beyond presentations and puzzles.

Conclusion

After the designer has looked at the learner, the learner’s environment, the performance goal, the gap analysis, etc., the performance might be whittled down to a problem type that the learner may need to solve.  The problem can range from providing performance feedback to evaluating equations, from delegating authority to applying statistics to improve production.   The type of program often suggests the treatment.  The fun of being a learning experience designer is in exploring strategies and applying them where appropriate in a cost-effective, time-saving manner.

In past articles, I have covered some of these strategies. You’re welcome to visit them.

ShortSims

Interactive Storytelling

Interactive Case Studies

Decision-Making Scenarios

Simulations

References

¹Jonassen, D.H. Toward a design theory of problem solving. ETR&D 48, 63–85 (2000). https://doi.org/10.1007/BF02300500

²In a 2021 survey held by Devlin Peck (n=615) (Instructional Designer Full Report 2021 | Devlin Peck), 44% of the respondents had 0 – 3 years’ experience.

³Aldrich, Clark (2020) ShortSims, CRC Press

 ⁴ Ambrose, S., Bridges, M., & Lovett, M. (2010). How learning works: 7 research-based principles for smart teaching. John Wiley and Sons.)